CellProfiler Analyst hangs at 99% when scoring all using RandomForest Classifier


#1

Hi CPA Help,

I’m new to CPA and I’m having some trouble when scoring all objects in my experiment.
I’m able to create a training set (2 bins: positive and negative), find rules (max 20) and score single images using RandomForest, but when I try to score all images, the progress bar gets to 99% and then seems to remain there indefinitely or until CPA stop responding entirely and has to be shut down. No error appears and the logger doesn’t contain any information as to what’s happening. My experiment contains approximately 16,000 images (25 fields/well in 4 plates) and I’m trying to group scores by plate and well. Is this simply too many images for a single computer using SQlite?
I’m using the latest nightly build on a Windows system.

Any help would be greatly appreciated!

Thanks,

Karla


#2

This seems like it might be related to a known issue that we are hoping is fixed in the master; unfortunately the builds have been broken for a while (but should be up again soon!) so the only way to test on your data if the fix worked would be to build from source.

Does it work if you try the Fast Gentle Boosting algorithm rather than Random Forest? That might be a good workaround for the short term if you don’t mind switching classifier algorithms.


#3

Thanks for the fast reply!
I tried Fast Gentle Boosting and had no problems at all so will make the switch for now.
If I need to go back to Random Forest I’ll give building from source a try.

Thanks so much for your help,

Karla


#4

Hi CellProfiler team,

I was just wondering if there has been any further progress on this issue?
We’re finding that Fast Gentle Boosting is the only algorithm that we can get to work at the moment. All of the other classifiers appear as if they’ll remain at 99% indefinitely. It’s especially weird because we initially had some success with the AdaBoost and Gradient Boosting Classifiers, but now we’re seeing the same issue with those as well. We’re using the latest Nightly (Windows), and are hoping to avoid having to build from source due to a lack of programming experience. Any idea if the Nightly build will be updated soon?

Thanks again for your help,

Karla


#5

Hi Karla,

As CPA is free and unfunded by any grants, it’s maintained on primarily a volunteer basis (and our current volunteer has been working on writing her PhD thesis at the moment); we’ve had some issues with our CP Windows builds as well, the consequence being that the CPA build issues aren’t fixed yet and likely won’t be until after we can find the solution in CP then apply it to CPA. If it’s absolutely critical to use another classifier in the meantime, I’m afraid source is your only option; I can absolutely understand that installing from source can be a pain (and kind of scary, at least it was for me!) when you’re not used to doing it, but the step-by-step instructions are pretty thorough. I’m sorry for the delay!


#6

No worries! Just wanted to check that I haven’t missed something, and to get an idea whether it’s worth taking the time to try installing from source if the new builds are going to be a while.

Thanks for your quick reply and keep up the awesome work!

Karla